Quantifying User Error in the Unsupervised Classification of Reservoirs

Natural resources such as water need to be prudently managed and accurately quantified to make informed decisions. Remotely sensed images collected by satellites can be used to estimate the surface area of water bodies such as reservoirs. However analyst bias can introduce uncertainty in area estimates. The purpose of this study was to quantify the extent of analyst bias introduced during image classification. Thirty LANDSAT 5 satellite images of six reservoirs were selected for this study. Pixels in each image were grouped into clusters using ISODATA algorithm that attempts to minimize the differences within groups and maximize differences between groups. Three analysts independent classified these images and assigned the clusters to water or non-water classes using their expertise. Classified images were combined to show the levels of agreement between users. Results indicate a high level of agreement between classified images, as well as spatial patterns in the distribution of user disagreement. Unsupervised classification methods are a viable option for estimating reservoir surface areas and user error generated by such methods can be minimized.